Artificial intelligence (AI)
Work more efficiently with a Swiss AI for your company.
For your artificial intelligence, we choose the model that suits your goal, connect it with your company’s knowledge and operate everything in our data centers in Switzerland.
100% Swiss
Your AI runs in our own data centers in the Zurich area, operated by Netstream.
Swiss law
Your data is subject to Swiss law, with no foreign cloud provider in between.
Training stays with you
You can train models with your data. However, your data does not flow back into public models that anyone can use.
Certified security
We operate our infrastructure in accordance with ISO/IEC 27001 and comply with the revised Swiss Data Protection Act (revDSG).
Save time
The AI supports your team with routine tasks and, depending on the goal, automates entire areas.
Your knowledge stays in-house
Internal documents, offers and customer data are only accessible to your company.
Assistant for your team
Your staff write, translate, summarise and research with an assistant that runs in Switzerland. They sign in with their existing company account, so no new passwords are needed. The interface knows the person’s role and only shows the models that are approved for their data.
- Single sign-on via your identity provider
- Model choice per confidentiality level
- Usage per person, department and month
Answers from your knowledge
The assistant accesses policies, regulations, contracts, minutes and knowledge bases and answers with source references. It only sees the documents that the person asking is allowed to open themselves.
- Permissions adopted from the Office environment, wikis and file storage
- Hybrid search combining full text and meaning, with reranking
- Index stays in Switzerland, embeddings from our own model
Integration into your applications
Agents with guardrails
Agents categorise tickets, reconcile invoices, prepare requests and check systems. Every action runs through a guardrail engine. What you have not allowed does not happen, and anything sensitive waits for approval.
- Mandates with purpose, scope and time window
- Approvals via chat or email
- Immutable audit log and kill switch
Dedicated environment
Architecture
One platform between your staff, the models and your systems.
Three building blocks that we developed ourselves and operate in our data centers in Switzerland. At the core are identity and the data boundary: every request runs with the person’s permissions, and rules determine which data reaches which model.
Chat interface
In the browser and on smartphones, sign-in via your identity provider.
Claude, ChatGPT, Copilot, Cursor
Your existing assistants access approved tools via MCP.
Software
OpenAI-compatible API for websites, business applications and workflows.
Every call runs with the person’s token, not via a service account
Netstream AI platform
LLM Gateway
- Identity from the Cortex token
- Egress policy per data classification
- Model routing and fallback
- Tracing according to OpenTelemetry GenAI
- Usage per person
Cortex · Agent Control Plane
- Roles and namespaces
- Mandates with purpose and time window
- Guardrail engine with five verdicts
- Approvals by humans
- Audit log
- Connectors with the person’s credentials
Retrieval
- Ingest from Office environments, wikis and file storage
- Permissions per document
- Hybrid search with reranking
- Eval harness with golden sets
Local inference
stays in-house- vLLM on our own GPU nodes
- Qwen, Apertus, Llama, Mistral, DeepSeek, gpt-oss
- Embeddings and rerankers
- Dedicated GPUs per customer on request
External inference
only by rule- Claude, GPT and others, only if the policy allows it for the respective classification
- Contractually without training
- Every transfer visible in the trace
Connectors, under the person’s identity
Office and communication
Mail, calendar, chat, document storage
Finance and ERP
Contacts, invoices, accounting
Ticketing and knowledge
Tickets, wikis, team communication
Infrastructure
Servers, container platforms, virtualisation, network
Business applications
REST, databases, custom connectors
For whom
How AI supports your industry.
AI pays off wherever confidential data is involved. The building blocks are the same, but they are used differently depending on the industry.
Healthcare
- Assistant
- Summarise reports and medical histories
- Knowledge
- Find internal guidelines with source references
- Agents
- Prepare documentation, completion after approval
Municipalities and public administration
- Assistant
- Draw up minutes
- Knowledge
- Search regulations
- Agents
- Prepare answers to enquiries
Law firms
- Assistant
- Prepare drafts
- Knowledge
- Research in your own files and contracts
- Agents
- Evaluate contracts and flag deviations
IT and software development
- Agents
- Analyse errors from the ticket system, fix as a pull request after approval
- Integration
- Integrate AI into your own applications
- Operations
- Put prototypes built with AI into operation with us
Fiduciaries, financial service providers, education and SMEs also benefit when the application suits the task. We are happy to work out together where AI relieves you the most.
Models, measured
We recommend the model that performs best on your cases.
An example from our eval harness: ticket triage, classification into five categories, twelve cases, all via the same chain. Local models run on an RTX PRO 6000 with 96 GB, external models via the LLM Gateway with an egress rule.
Measurement example: ticket triage
Twelve cases, five categories
| Model | Operation | Accuracy | Macro F1 | parallel @ 8k |
|---|---|---|---|---|
| Qwen2.5-7B-Instruct | local | 83.3 % | 0.825 | 160× |
| swiss-ai Apertus-8B | local | 75.0 % | 0.748 | 70× |
| Qwen2.5-14B-Instruct | local | 83.3 % | 0.900 | 37× |
| Qwen2.5-32B-Instruct | local | 91.7 % | 0.927 | 11× |
| Claude Haiku 4.5 | external, by policy | 91.7 % | 0.927 | – |
| Claude Sonnet 5 | external, by policy | 100 % | 1.000 | – |
For this task, the local 32B model achieves the same scores as the external Haiku. The data does not have to leave the company.
Capacity instead of gut feeling
What matters is how many staff can work with a model at the same time. Qwen2.5-32B needs 2.0 GiB per request with 8k context.
KV cache (GiB) ≈ 0.9 × VRAM − 1.03 × weights − 1
| GPU | VRAM | KV cache | Sessions |
|---|---|---|---|
| RTX PRO 6000 | 96 GB | 22.5 GiB | 11 |
| H200 | 141 GB | ~60 GiB | ~30 |
| B200 | 180 GB | ~95 GiB | ~47 |
| B300 | 288 GB | ~193 GiB | ~96 |
First row measured, the others extrapolated from the measurement. Before a purchase, we measure on rented hardware, with your cases.
From use case to solution
We find the right AI with you.
Before we recommend a model, we measure it against your cases. That way, you can plan with actual project costs.
- 01
Conversation
What goal are you pursuing and what data is involved?
- 02
Model selection workshop
You bring 30 real cases. We measure local and external models against your task.
- 03
Solution design
Model, size, policy and capacity, with a substantiated recommendation and actual project costs.
Operating model: shared platform or dedicated environment with reserved GPUs, your own model and your own namespace.
- 04
Introduction
We set up the platform, connect your systems and support your team as they get started. We then continuously improve the solution based on your experience.
- 05
Operation
In our data centers in Switzerland, with monitoring and updates.
FAQ
Frequently asked questions
Where does the AI run?
Does my data flow into public models?
Which models are available?
How do staff sign in?
Can I integrate AI into my own applications?
Is there a dedicated environment?
What availability applies?
Unless otherwise agreed, our IT and cloud services have an availability of 99.9 % per calendar year. We report incidents on status.netstream.ch.
Docs: Availability of our servicesWhat happens to my data when the contract ends?
Your data belongs to you. Back it up or export it before the contract ends. We then delete it within 30 days and confirm the deletion on request.
Docs: Data return and deletion at the end of the contractWho is responsible for what?
Netstream operates the infrastructure, data center and platform. You are responsible for the configuration, your users and backing up your data.
Docs: Shared responsibility